Optimal bipartite scorecards

摘要

In many retail banking applications, scorecards used for assessing creditworthiness must be simple and interpretable. For this reason, the industry has favoured logistic regression models based on categorised variables. In this paper we describe an extension of such models based on an optimal partition of the applicant population into two subgroups, with categorised logistic models being built in each part. Such bipartite models have the merits of yielding improved predictive accuracy, while retaining interpretive simplicity. They have been used in the industry before, but only in an ad hoc way, with no effort being made to find the optimal division. Some examples and properties of the resulting models are described.